3D Cartoon Face Generation with Controllable Expressions from a Single GAN Image

  • 2022-07-29 02:06:21
  • Hao Wang, Guosheng Lin, Steven C. H. Hoi, Chunyan Miao
  • 43

Abstract

In this paper, we investigate an open research task of generating 3D cartoonface shapes from single 2D GAN generated human faces and without 3Dsupervision, where we can also manipulate the facial expressions of the 3Dshapes. To this end, we discover the semantic meanings of StyleGAN latentspace, such that we are able to produce face images of various expressions,poses, and lighting by controlling the latent codes. Specifically, we firstfinetune the pretrained StyleGAN face model on the cartoon datasets. By feedingthe same latent codes to face and cartoon generation models, we aim to realizethe translation from 2D human face images to cartoon styled avatars. We thendiscover semantic directions of the GAN latent space, in an attempt to changethe facial expressions while preserving the original identity. As we do nothave any 3D annotations for cartoon faces, we manipulate the latent codes togenerate images with different poses and lighting, such that we can reconstructthe 3D cartoon face shapes. We validate the efficacy of our method on threecartoon datasets qualitatively and quantitatively.

 

Quick Read (beta)

loading the full paper ...